FieldFix — Offline AI Repair Assistant for Field Technicians

FieldFix was built for the environments where repair problems actually happen: agricultural fields, industrial facilities, remote construction sites — places with no cloud connectivity and no time to search through manuals. A technician speaks or types a symptom, and FieldFix returns structured repair guidance: ranked probable causes, step-by-step instructions, required tools, and safety stop conditions.

The AI backbone is Gemma 3 4B running through Ollama with Metal GPU acceleration on the host laptop. The model never touches the cloud — all inference is local. A RAG layer (ChromaDB with all-MiniLM-L6-v2 embeddings, 295 semantic chunks) grounds the model's responses in 37 expert-written repair documents spanning Robotics, Electronics, Emergency Equipment, Household systems, and Safety Guides.

The multi-agent pipeline processes each symptom through five specialized agents: a diagnosis agent identifies probable causes, a cause ranker orders them by likelihood, a repair planner generates RAG-augmented step-by-step instructions, a question agent surfaces clarifying questions, and a verification agent produces stop conditions and prevention tips.

Safety is handled deterministically — before any AI agent sees a query, a rules-based guardrail layer classifies it against 9 hard-stop categories (gas leaks, electrical fire, lithium battery failure, high-voltage exposure, etc.). Hard stops return an immediate warning with no AI processing. This design choice — no model for safety decisions — is intentional.

A per-device SQLite history store lets the system recall previous repairs on the same device, improving diagnostic accuracy for recurring issues. The frontend is a Next.js app served over the local network, accessible from iOS Safari on any phone connected to the same WiFi.